AAPS PharmSciTech ( # 2018)

DOI: 10.1208/s12249-018-1173-2

Research Article

Comparative Study for Optimization of Pharmaceutical Self-Emulsifying

Pre-concentrate by Design of Experiment and Artificial Neural Network

Kinjal J. Parikh1 and Krutika K. Sawant1,2

Received 23 May 2018; accepted 31 August 2018

Abstract. The present investigation aimed to optimize the critical parameters affecting the
globule size of self-emulsifying drug delivery system. Based on preliminary screening, three
critical parameters, viz., amount of oil, surfactant, and co-surfactant were found to affect the
globule size. I-optimal mixture design and Artificial Neural Network (ANN) were used to
optimize the formulation with respect to minimum globule size. Comparative study was
carried out to identify which optimization technique gave better predictability for the selected
output parameter. R-value and MSE values were taken into consideration for comparison of
both techniques. Using Response Surface Methodology-based I-optimal mixture design
approach, the R2 value was found to be 0.9867, whereas with ANN technique, it was found to
be 0.99548. The predicted size for the optimized batch by I-optimal design was 122.377 nm,
whereas by ANN, it was 119.6783 nm against the actual obtained size of 118.2 ± 2.3 nm. This
analysis indicated superior predictability of output for given input variables by ANN as
compared to model-dependent DoE I-optimal design approach.
KEY WORDS: self-emulsifying pre-concentrate; formulation optimization; I-optimal design; multiple
regression; artificial neural network.

INTRODUCTION nano-lipid carriers (8,9). However, the techniques for formu-

Among various approaches for enhancing dissolution of
poorly soluble BCS class II drugs, self-emulsifying pre-
concentrates (SEPCs) have proven to be more effective (1).
This pre-concentrate is made up of oil, surfactant, and co-
surfactant. After dispersion with media in the g.i. tract, it
forms globules of < 200 nm size (2). These globules, based on
the chain length of oil, enter into the lymphatic circulation
bypassing first pass metabolism in the liver. The nano-size of
the globules can enormously enhance the surface area and
increase the absorption rate from the intestine in comparison
to drug suspension (3,4). Increased drug solubilization in the
intestinal milieu, intestinal lymphatic transport, and avoid-
ance of first pass hepatic metabolism leads to increase in
bioavailability (5).
Iloperidone (ILO), a weakly basic atypical antipsychotic,
is a second-generation drug which binds to 5-HT2a receptor
with higher affinity in the mesocortical system (6). The drug is
practically insoluble in water and undergoes first pass
metabolism, resulting in only 36% absolute oral bioavailabil-
ity (7). There are only two reports till date for enhancement
of dissolution rate of Iloperidone using microspheres and
1
Faculty of Pharmacy, The M S University of Baroda, Kalabhavan,
Vadodara, Gujarat 390 001, India.
2
To whom correspondence should be addressed. (e–mail:
dr_krutikasawant@yahoo.co.in)
lation development used therein are highly energy consuming
and involve many variables which could affect the final
formulation characteristics. So far, there are no reports for
dissolution enhancement of ILO by SEPCs. Hence, the
present investigation was targeted to develop SEPCs of ILO
which upon dilution can produce o/w microemulsion that can
provide large surface area for absorption to increase its
bioavailability.
As the preparation of SEPCs is a multivariate process,
thus, the final characteristics of the formulation are affected
by various formulation parameters. In such cases, conven-
tional one variable at a time (OVAT) approach cannot be
used as it requires large number of experiments to identify
the main effect parameters (10). Hence, design of experiment
(DoE)-based statistical technique is recommended for screen-
ing and optimization steps.
Response surface methodology (RSM) is widely used for
optimization of drug formulations (11). It differs from the
screening design in terms of modeling of parameters, as it
enables fitting of not only linear but also non-linear models
such as cubic and quadratic. Mapping of the response by
contour plot and 3D response curve enables selection of
optimum solution (12). I-optimal design, a prediction-
oriented mixture design, is used herein for formulation
optimization. It can give good prediction over design space,
as against D-optimality which is related to covariance of
matrix of parameter estimates (13). Here, coordinate

1530-9932/18/0000-0001/0 # 2018 American Association of Pharmaceutical Scientists


exchange optimal search method for I-optimality criteria was
used to build an optimal design, as this algorithm for
coordinate exchange gives optimal design (13,14).
Despite the advantages of DoE–RSM, literature review
suggests that DoE-based prediction of responses constructed
using polynomial equation is often inadequate to describe the
complex interrelationship between variables/factors and re-
sponses (15). To overcome this, an alternative approach of
Artificial Neural Network (ANN) has been suggested.
ANN, also known as neuro-computing or parallel
distributed processing, is a branch of artificial intelligence
(AI) which is inspired by biological nervous system (16).
ANN is designed to simulate the human brain function which
are model-independent computational paradigms (17). It can
be applied to identify non-linear relationship between causal
factors and their responses by pre-decided iteration-based
data training. The preliminary trials for formulation develop-
ment are used as input information for training the neural
network, and once trained, this network is used to forecast
output by simulation of the trained network.
In this study, we have optimized the high-risk formula-
tion variables affecting SEPCs by employing two different
statistical tools, I-optimal design and ANN, and have
compared these techniques for the prediction of target output
data, i.e., minimal globule size.
MATERIALS AND METHODS
Materials
Iloperidone was received as a gift sample from Alembic
Pharmaceuticals Ltd., India. Capmul MCM C8 was received
as a gift sample from Abitec Corporation, USA. Oleic acid
was purchased from Loba Chemie., India. Cremophor EL
was purchased from Sigma-Aldrich, India. Transcutol HP was
received as a gift sample from Gattefosse, India. All other
chemicals used were of analytical grade.
Pseudo-Ternary Phase Diagram
After screening of oil, based on its solubilization capacity of
the drug ILO, surfactant and co-surfactant were screened based on
emulsification efficiency measured in terms of %Transmittance
(18,19). Capmul MCM C8/Oleic acid (1:2) was selected as oil
phase, and Cremophor EL and Transcutol HP were selected as
surfactant and co-surfactant, respectively. Pseudo-ternary phase
diagrams were constructed using different proportions of oil to
surfactant/co-surfactant (Smix) ratio at room temperature. For the
given weight ratio, the total of oil, Smix, and water always added to
100%. The weight ratio of oil/Smix was varied from 1:9 to 9:1. Two
hundred milligrams of each weight ratio was titrated slowly with
distilled water to allow equilibrium to produce fine microemulsion.
Preparation of SEPCs
Design of Experiment by I-Optimal Mixture Design
I-optimal design was used as a statistical tool to quantify the
relationship between the critical formulation factors and mea-
sured response for optimization. As I-optimal design is a
mixture design, the total of all the formulation variables made
Parikh and Sawant
up to 100% (20). Based on the screening studies, the following
parameters were identified as key factors for formulation of self-
emulsifying pre-concentrate: (a) amount of oil, (b) amount of
surfactant, and (c) amount of co-surfactant.
A 16-run, three factors, two-level I-optimal mixture
design was employed to study the effect of formulation
(independent) variables on globule size (dependent variable).
Preliminary trials were performed to select the discrete levels
of the independent variables. Table I summarizes the
independent and dependent variables evaluated and the goal
set for the response.
Based on these input ranges for independent variables
and goal for dependent variable, a matrix was created by the
Design expert 10.0 stat-ease software. Polynomial model was
generated for the response, i.e., globule size. Model coeffi-
cients were generated after conversion of actual component
proportion to L-pseudo-level using Eq. 1.
Component weight fraction−Li
L−pseudo ¼ ð1Þ
1−L
Li is the lower constraint of component weight fraction
and L is the sum of lower constraint of components’ weight
fraction (21).
Component weight fraction was calculated as:
Component amount
Component weight fraction ¼ ð2Þ
Sum of components0 amount
Best fit model was selected based on the R2 value. After
exclusion of statistically insignificant terms by ANOVA,
contour plot and fraction of design space (FDS) plots were
generated (22).
Artificial Neural Network
ANN and MATLAB Software. ANN has recently re-
ceived wide attention in the field of formulation development.
ANN is composed of processing units termed as Bneurons^ which
are interconnected forming a network (23). Here, we have used
multilayer perceptron (MLP), which is organized as a set of
interconnected input, hidden, and output layers (24). The number
of input layers is the number of input variables, i.e., 3 in this
network. The number of hidden layers is adjusted to decrease the
error between experimental target and simulated output values
(25). The number of output layers is the number of dependent
variables. When the neural network is provided with the input
data, the input layer propagates the weighted data to the hidden
Table I. Variables for I-Optimal Design for Optimization of SEPCs
Independent variables Low level High level
Oil-Capmul MCM C8/Oleic acid:: 1:2 (A) 8 mg 14 mg
Surfactant—Cremophor EL (B) 56 mg 65 mg
Co-surfactant—Transcutol HP (C) 25 mg 35 mg
Dependent variables Goal Importance
Globule size Minimize +++
Artificial Neural Network for formulation optimization
Fig. 1. MLP feedforward back propagation network
layer. After selection of the bias from this layer, output response
is provided by using transfer function. Every time, during learning
process, there is adjustment of the interconnection weights
between layers (16). After such training, the network predicts
outputs for new set of input data, which is termed as network
simulation. This defines the generalization ability of the created
network. For MLP network formation by feed forward back
propagation algorithm, MATLAB R2015b uses 70% of input
data for training, 15% data for validation, and remaining 15%
data for testing purpose (26).
Creating the MLP Feed Forward Back Propagation
Network. This MLP network is created for supervised
learning problems, meaning that there is training set of input
and output so that the network learns interdependency
between them. During this, the weights and biases are
adapted to optimal value, which can be evaluated by its
performance and mean square error (MSE) values. The
activation (transfer) function for MLP was chosen to be
logistic sigmoid (25). The hidden layer contains 10 neurons.
Training algorithm was chosen as Levenberg-Marquardt. To
avoid overtraining of the network, 1000 iterations were fixed.
Training continued till improvement in MSE value was seen.
Training auto-stopped when generalization improvement did
not occur, i.e., when MSE value did not decrease.
Fig. 2. Pseudo-ternary phase diagram of Oil [Capmul MCM C8/Oleic acid:: 1:2] + Smix
[Cremophor EL/Transcutol HP:: 3:1) + water
The input datasets were divided as training (70%),
validation (15%), and testing (15%). The training data were
presented to network during training and the network
adjusted according to its error. The validation data were used
to measure the network generalization. Based on this, the
network stopped training as the generalization stopped
improving. Testing data were an independent measure of
network performance (26).
Figure 1 shows the MLP feedforward back propagation
network created using MATLAB 2015b.
Network Training. In this learning and training process,
weights and bias of the network changes constantly. This
leads to different output due to different weights. Lastly, one
can achieve appropriate network after running number of
iterations that minimize MSE.
The algorithm for the network was feed forward back
propagation. During the forward pass, the predicted
outputs corresponding to given inputs were evaluated as
per Eq. 3.
X ¼ f ðsÞ ¼ BμðAs þ aÞ þ b ð3Þ
where s is a vector for outputs. A and a are the matrix of
weights and bias of the first layer, respectively. B and b
 Parikh and Sawant

Table II. Three-Component I-Optimal Design for Optimization of SEPCs

Run Space type Component A Component B Component C Response 1

Oil (mg) Surfactant (mg) Co-surfactant (mg) Size (nm) ± SD

1 Interior 11.0322 62.1051 26.8627 118.70 ± 2.12

2 Edge 13.725 56 30.275 194.30 ± 1.93
3 Edge 8 64.707 27.293 24.33 ± 2.14
4 Interior 11.0621 59.6453 29.2926 127.60 ± 3.01
5 Edge 8 62.363 29.637 33.26 ± 2.18
6 Interior 11.0621 59.6453 29.2926 126.50 ± 3.90
7 Vertex 10 65 25 99.91 ± 1.74
8 Vertex 9 56 35 69.09 ± 1.61
9 Interior 10.8247 57.1891 31.9681 133.70 ± 2.56
10 Vertex 14 61 25 174.30 ± 3.63
11 Edge 14 58.3932 27.6068 195.90 ± 1.39
12 Vertex 9 56 35 53.80 ± 2.33
13 Edge 8 59.5448 32.4552 12.34 ± 1.66
14 Vertex 10 65 25 88.95 ± 2.48
15 Interior 11.0621 59.6453 29.2926 124.10 ± 3.53
16 Edge 13.725 56 30.275 194.90 ± 2.37

are the matrix of weight and bias for the second layer,
respectively. The function μ denotes elementwise
nonlinearity.
Reverse transmission was used to transmit error layer
by layer. After adjustment of the connection weights
among neurons, MSE value was decreased leading to best
performance of the network to generate output.
Globule Size Analysis
The SEPC was diluted to 10 mL with water. Visual
observations were done to check appearance, transparency,
phase separation, and precipitation of drug. Average globule
size was measured using Zetasizer Nano ZS90 (Malvern
Instruments Ltd., UK) which is based on the principle of

Fig. 3. Fraction of design space (FDS) graph

Artificial Neural Network for formulation optimization

Table III. Model Selection Based on the R2 Value for I-Optimal small chain fatty acid/triglycerides such as Capmul MCM C8
Design showed spontaneous emulsification upon dilution. So, combi-
nation of the two oils in the optimized ratio of Capmul MCM
C8/Oleic acid:: 1:2 was chosen as oil component. This
Model Adjusted PRESS optimization was done using OVAT approach for high
R2 value value solubilization and spontaneous emulsification (29).
Cremophor EL and Transcutol HP were selected as surfac-
Linear 0.9575 3269.70
Quadratic 0.9867 1331.28 Suggested
tant and co-surfactant, respectively, based on the %Transmit-
Special cubic 0.9854 1883.23 tance taken at 638.2 nm (19). The phase diagram for these
Cubic 0.9904 8.386E + components was prepared by varying the range of Oil/Smix
005 from 1:9 to 9:1 (Fig. 2), keeping the ratio of surfactant/co-
Special quartic vs cubic 0.9869 13,172.12 surfactant at 3:1 (30).
Quartic vs cubic 0.9901 – Aliased
Quartic vs sp. quartic 0.9901 – Aliased
Batch Preparation as per I-Optimal Design

In this computer-generated I-optimal design, the total of

dynamic light scattering (1,27). A standard 633-nm laser fitted
with 90° scattering optics was used. Temperature was set to
25 °C, and sample equilibration time was kept 120 s. The size
was measured by Brownian motion of the globules in water as
the medium by photon correlation spectroscopy.
the mixture was 100%, i.e., the fractions of each component
A, B, and C summed up to 1 in the final formulation. Sixteen
batches were prepared as per I-optimal design generated by
design expert 10.0.0 software (31). The response in terms of
size (nm) has been given in Table II (32).

RESULTS AND DISCUSSION Analysis of Mixture Design

Pseudo-Ternary Phase Diagram
Selection of oil for the preparation of SEPCs was based
on solubility study (18,28). Even though Iloperidone showed
high solubility in long chain fatty acid, oleic acid, it could not
spontaneously form nano-sized emulsion, whereas use of
The space type of design matrix was generated by the
computer-based mixture design algorithm as shown in
Table II. For each run, build type was also generated.
Build type indicates whether the run is model point, lack
of fit point, or replicate point from the generated matrix
(22). The model points are used for estimation of all

Fig. 4. Actual vs. predicted graph for response–size (nm)

 Parikh and Sawant

Fig. 5. 2D contour plot for droplet size

coefficients, lack of fit points test how well the model The condition number of this matrix was 175.677. This
represents the actual behavior and the replicate points value indicates moderate to strong multicollinearity, i.e., the
estimate the pure error. predicted variable (size) in the multiple regression model

Fig. 6. 3D response plot for droplet size

Artificial Neural Network for formulation optimization

Table IV. Optimized Formulation Composition

Component A (oil) Component B (surfactant) Component C (co-surfactant) Size (nm) Desirability

Suggested 11.147 63.853 25.000 122.377 0.458

Prepared 11.147 63.853 25.000 118.2 ± 2.3 –
%Prediction error 3.53%

(quadratic model) is highly correlated with the predictor The effect of different proportion of components on size
variables (components A, B, and C) (33). could be explained by Eq. 4.
Fraction of design space (FDS) graph was generated at
process sigma, s = 1 and α risk level of 0.05 as shown in Fig. 3. Globule size ðnmÞ ¼ 115:41 þ 2657:54  A−365:68
It displays the area/volume of the design space having a mean
standard error less than or equal to a specified value. The  B−215:99  C þ 498:61  AB
ratio of the area/volume to total area/volume is the fraction of þ 10696:27  AC−1166:17  BC ð4Þ
design space. This graph as shown in Fig. 3 indicated that
50% of the design space had a relative standard error of less
From the quadratic equation, it can be concluded that
than 0.5 suggesting that predictability using this design space
there was complex effect of all components on size. As the
was very less (22).
value of coefficient for component A was highest, component
Based on the statistical analysis by ANOVA, models
A (oil) had most prominent effect on size as compared to B
were generated for interpretation of size from I-optimal
and C. By increasing amount of component A, there will be
design as shown in Table III.
increase in size, whereas by increasing amount of B and C in
Based on the R2 value of 0.9867, quadratic model was chosen
combination, there will be decrease in size, as the coefficient
for optimization of size. Even though the adjusted R2 value was
for BC showed negative effect on size. The value for
high for quartic model, it was aliased due to lack of orthogonality.
coefficient for BC is much lower as compared to coefficient
The PRESS (predicted residual error sum of squares) statistics also
terms of B and C, indicating that the amount of B and C
indicated that the selected model, quadratic, was precise. This was
chosen for SPECs based on phase ternary diagram is very
confirmed from the graph of actual vs. predicted size as shown in
critical for globule size of oil.
Fig. 4. The MSE for the quadratic model was 48.94 indicated low
The globule is oil (A) droplet surrounded by surfactant
predictability using this model.
(B) and co-surfactant (C) (34). Without oil, surfactant and co-
surfactants form micelles which are having much lower size
Fitting of the Model and Analysis than oil globules. This indicates the reason behind high
coefficient for A and negative coefficient values for B and C.
The relationship between experimentally obtained glob- Presence of co-surfactant only with oil is not efficiently
ule size and formulation variables was determined in terms of able to decrease size (higher coefficient value of AC),
actual components used as input variables. whereas surfactant can decrease size of oil globules (lower

Fig. 7. Globule size of the optimized formulation

 Parikh and Sawant

Fig. 8. Architecture of artificial neural network for optimization of globule size

Fig. 9. Regression of dataset using MATLAB. R value for (a) training dataset, (b) validation dataset, (c) test dataset, and
(d) overall
Artificial Neural Network for formulation optimization
Fig. 10. Effect of input variables on output. a Effect of oil and surfactant on size.
b Effect of oil and co-surfactant on size. c Effect of surfactant and co-surfactant on
size
Fig. 11. Relative contribution of each ingredient on size
value of AB coefficient). The combined effect of surfactant
and co-surfactant decreases size of oil globules (negative
value of BC coefficient).
The contour plot is shown in Fig. 5, and the 3D response
plot for the same is shown in Fig. 6. The 3D plot clearly indicates
non-linear interrelationship of 3 input variables on output size.
Optimized Formulation by Desirability Function
The optimized formulation, having least globule size, was
selected based on the desirability value from the suggested
results. The formulation composition having highest desir-
ability of 0.458 was formulated in triplicate, and globule size
was measured (Table IV). Due to applied constraints on input
variables and output variables, the highest desirability value
was 0.458. Removing the constraints increased desirability
value, but the obtained formulation in such cases did not
meet the quality target product profile. Figure 7 shows the
globule size of the optimized formulation. The Polydispersity
Index (PDI) was found to be 0.285. PDI is calculated from

cumulant analysis of measured intensity. This is autocorrela-
tion function which describes width of the Gaussian distribu-
tion of the size (35). PDI ranges from 0 to 1. Lesser value of
PDI indicated mono-dispersed formulation.
Artificial Neural Network
Structure of ANN
ANN feed forward back propagation framework was
fitted to the data of I-optimal design. The data set was divided
into training (12 sets), validation (2 sets), and testing data (2
sets). The training of the network by Levenberg-Marquardt
topology minimized the error sum of square for the training
dataset to give high R value. During the training, the weights
and biases of the network were adjusted iteratively to
minimize the network performance function. The MSE is
the default performance function for the generated network.
The Learning function was gradient descent with momentum
(learngdm), because the momentum allows the network to
ignore small features in the error surface. The number of
training cycles was on the basis of MSE of the validation
dataset (16,21). Figure 8 shows the architecture of ANN.
Regression
Regression of the neural network dataset showed high R
value. Overall combined value of the regression coefficient
was found to be 0.99548 (Fig. 9) which was higher in
comparison to regression coefficient value of 0.9867 of the
quadratic model of I-optimal design. This indicated that ANN
model was trained to give better results in comparison to the
I-optimal design of experiment (26). The MSE of the ANN
model was 0.014 indicated that ANN had better predictability
to measure the quality of estimator, i.e., size in this research.
Value of MSE near to zero is considered to be good.
The 3D surface plots for the three input data against the
output are shown in Fig. 10. The 3D graphs suggest that by
increasing the oil content, the size increases, whereas by
increasing the concentration of surfactant and co-surfactant,
the size decreases. The relative contributory effect of each
input on the output is shown in Fig. 11. It also shows that the
output (size) is mainly dependent on the level of oil, with
54% relative contribution, whereas surfactant and co-
surfactant have equal contribution of 23% each.
Simulation of the ANN Network
For the simulation of the generated network, the same
dataset as the one suggested by the I-optimal design was used
as input for ANN (36). The amount of oil, surfactant, and co-
surfactant was 11.147, 63.859, and 25.000 mg, respectively.
The output size generated for the simulation dataset was
119.6783 nm. The batch prepared using the same composition
gave the actual size of 118.2 ± 2.3 nm. The %prediction error
for the output was found to be 1.25%.
Comparison of I-Optimal Design and ANN
The R value for the I-optimal design-based quadratic
model was 0.9867, whereas for ANN, it was 0.99548. The
Parikh and Sawant
MSE for quadratic model was 48.94, whereas for ANN, it was
only 0.014. The size predicted by the I-optimal design was
122.377 nm, whereas by ANN, it was 119.6783 nm. The actual
size was found to be 118.2 ± 2.3 nm. %Prediction error was
found to be 3.53% by DoE, whereas by ANN, it was 1.25%.
All this indicated better predictability power of the ANN
over I-optimal design (37). Moreover, the error for ANN was
less as compared to I-optimal design, indicating its better
accuracy in prediction of the size (38,39).
CONCLUSION
In this study, the authors report the preparation and
successful optimization of Iloperidone self-emulsifying pre-
concentrate by I-optimal design and ANN approaches.
Contour plot and 3D graphs generated by both I-optimal
design and ANN showed interaction between input and
output data. The relative contribution of each ingredient on
size generated by the ANN indicated that oil showed the
highest effect on the output. The developed and validated
ANN model showed that higher predictability and accuracy
as compared to I-optimal design-based quadratic model based
on low MSE value were low and high R value were high.
The outcome of this study stresses the importance of
judicious choice of the optimization tool for incorporating
desired properties in a formulation.
ACKNOWLEDGEMENTS
The authors would like to acknowledge the support of
Chirag Makvana and Ankit Desai for their help in coding
evaluation. The authors would also like to thank Alembic
Pharmaceuticals Ltd., Vadodara, India, for providing
Iloperidone as gift sample.
COMPLIANCE WITH ETHICAL STANDARDS
Conflict of Interest The authors declare that they have no conflict
of interest.
REFERENCES
1. Gupta S, Chavhan S, Sawant KK. Self-nanoemulsifying drug
delivery system for adefovir dipivoxil: design, characterization,
in vitro and ex vivo evaluation. Colloids Surf A Physicochem
Eng Asp. 2011;392(1):145–55.
2. Singh AK, Chaurasiya A, Singh M, Upadhyay SC, Mukherjee
R, Khar RK. Exemestane loaded self-microemulsifying drug
delivery system (SMEDDS): development and optimization.
AAPS PharmSciTech. 2008;9(2):628–34.
3. Bandyopadhyay S, Katare O, Singh B. Optimized self nano-
emulsifying systems of ezetimibe with enhanced bioavailability
potential using long chain and medium chain triglycerides.
Colloids Surf B: Biointerfaces. 2012;100:50–61.
4. Sawant KK, Patel MH, Patel K. Cefdinir nanosuspension for
improved oral bioavailability by media milling technique:
formulation, characterization and in vitro–in vivo evaluations.
Drug Dev Ind Pharm. 2016;42(5):758–68.
5. Porter CJH, Trevaskis NL, Charman WN. Lipids and lipid-
based formulations: optimizing the oral delivery of lipophilic
Artificial Neural Network for formulation optimization
drugs. Nat Rev Drug Discov. 2007;6:231–48. https://doi.org/
10.1038/nrd2197.
6. Gray J, Roth BL. The pipeline and future of drug development
in schizophrenia. Mol Psychiatry. 2007;12(10):904–22.
7. Zhang T, Yang Y, Wang H, Sun F, Zhao X, Jia J, et al. Using
dissolution and pharmacokinetics studies of crystal form to
optimize the original iloperidone. Cryst Growth Des.
2013;13(12):5261–6.
8. Ige PP, Agrawal K, Patil U. Enhanced in vitro dissolution of
Iloperidone using Caesalpinia Pulcherrima mucoadhesive mi-
crospheres. Beni-Suef Univ J Appl Sci. 2015;4(1):26–32.
9. Mandpe L, Kyadarkunte A, Pokharkar V. Assessment of novel
iloperidone- and idebenone-loaded nanostructured lipid car-
riers: brain targeting efficiency and neuroprotective potential.
Ther Deliv. 2013;4(11):1365–83.
10. Singh B, Kapil R, Nandi M, Ahuja N. Developing oral drug
delivery systems using formulation by design: vital precepts,
retrospect and prospects. Expert Opin Drug Deliv.
2011;8(10):1341–60.
11. Singh B, Bhatowa R, Tripathi CB, Kapil R. Developing micro-/
nanoparticulate drug delivery systems using design of experi-
ments. Int J Pharm Investig. 2011;1(2):75. https://doi.org/
10.4103/2230-973X.82395.
12. Eriksson L, Johansson F, Kettaneh-Wold N, Wikström C, Wold
S. Analysis of factorial designs (Level 1) in design of experi-
ments: principles and applications. Stockholm: MKS Umetrics
AB; 2008. p. 85–92.
13. Jones B, Goos P. I-optimal versus D-optimal split-plot response
surface designs. J Qual Technol. 2012;44(2):85–101. https://
doi.org/10.1080/00224065.2012.11917886.
14. Meyer RK, Nachtsheim CJ. The coordinate-exchange algorithm
for constructing exact optimal experimental designs.
Technometrics. 1995;37(1):60–9.
15. Desai KM, Survase SA, Saudagar PS, Lele S, Singhal RS.
Comparison of artificial neural network (ANN) and response
surface methodology (RSM) in fermentation media optimiza-
tion: case study of fermentative production of scleroglucan.
Biochem Eng J. 2008;41(3):266–73.
16. Agatonovic-Kustrin S, Beresford R. Basic concepts of artificial
neural network (ANN) modeling and its application in pharma-
ceutical research. J Pharm Biomed Anal. 2000;22(5):717–27.
17. Basheer I, Hajmeer M. Artificial neural networks: fundamen-
tals, computing, design, and application. J Microbiol Methods.
2000;43(1):3–31.
18. Date AA, Nagarsenker M. Design and evaluation of self-
nanoemulsifying drug delivery systems (SNEDDS) for
cefpodoxime proxetil. Int J Pharm. 2007;329(1):166–72.
19. Yavas O, Do N, Tam AC, Leung PT, Leung WP, Park HK, et al.
Temperature dependence of optical properties for amorphous
silicon at wavelengths of 632.8 and 752 nm. Opt Lett.
1993;18(7):540–2. https://doi.org/10.1364/OL.18.000540.
20. Holm R, Jensen I, Sonnergaard J. Optimization of self-
microemulsifying drug delivery systems (SMEDDS) using a D-
optimal design and the desirability function. Drug Dev Ind
Pharm. 2006;32(9):1025–32.
21. Medarević DP, Kleinebudde P, Djuriš J, Djurić Z, Ibrić S.
Combined application of mixture experimental design and
artificial neural networks in the solid dispersion development.
Drug Dev Ind Pharm. 2016;42(3):389–402.
22. Myers RH, Montgomery DC, Anderson-Cook CM. Two-level
factorial designs in response surface methodology: process and
product optimization using designed experiments. Hoboken:
John Wiley & Sons; 2016. p. 81–160.
23. Anand P, Siva Prasad B, Venkateswarlu C. Modeling and
optimization of a pharmaceutical formulation system using
radial basis function network. Int J Neural Syst.
2009;19(02):127–36.
24. Dayhoff JE, DeLeo JM. Artificial neural networks. Cancer.
2001;91(S8):1615–35.
25. Svozil D, Kvasnicka V, Pospichal J. Introduction to multi-layer
feed-forward neural networks. Chemom Intell Lab Syst.
1997;39(1):43–62.
26. Hashad RA, Ishak RA, Fahmy S, Mansour S, Geneidi AS.
Chitosan-tripolyphosphate nanoparticles: optimization of for-
mulation parameters for improving process yield at a novel pH
using artificial neural networks. Int J Biol Macromol.
2016;86:50–8.
27. Patel D, Sawant KK. Self micro-emulsifying drug delivery
system: formulation development and biopharmaceutical evalu-
ation of lipophilic drugs. Curr Drug Deliv. 2009;6(4):419–24.
28. Patel D, Sawant KK. Oral bioavailability enhancement of
acyclovir by self-microemulsifying drug delivery systems
(SMEDDS). Drug Dev Ind Pharm. 2007;33(12):1318–26.
29. Patel K, Sarma V, Vavia P. Design and evaluation of
Lumefantrine–oleic acid self nanoemulsifying ionic complex
for enhanced dissolution. DARU. 2013;21(1):1.
30. Kumar M, Pathak K, Misra A. Formulation and characteriza-
tion of nanoemulsion-based drug delivery system of risperidone.
Drug Dev Ind Pharm. 2009;35(4):387–95.
31. Sawant K, Pandey A, Patel S. Aripiprazole loaded poly
(caprolactone) nanoparticles: optimization and in vivo pharma-
cokinetics. Mater Sci Eng C. 2016;66:230–43.
32. Thakkar H, Nangesh J, Parmar M, Patel D. Formulation and
characterization of lipid-based drug delivery system of
raloxifene-microemulsion and self-microemulsifying drug deliv-
ery system. J Pharm Bioallied Sci. 2011;3(3):442.
33. Paul R. Multicollinearity: causes, effects and remedies. Indian
Agric Stat Res Inst. 2008;4:2005–29.
34. Hathout RM, Woodman TJ. Applications of NMR in the
characterization of pharmaceutical microemulsions. J Control
Release. 2012;161(1):62–72.
35. Dynamic light scattering, Common terms defined. Inform White
Paper - Malvern Instruments Limited. 2011:1–6.
36. Alam FM, McNaught KR, Ringrose TJ. A comparison of
experimental designs in the development of a neural network
simulation metamodel. Simul Model Pract Theory.
2004;12(7):559–78.
37. Arulsudar N, Subramanian N, Murthy R. Comparison of
artificial neural network and multiple linear regression in the
optimization of formulation parameters of leuprolide acetate
loaded liposomes. J Pharm Pharm Sci. 2005;8(2):243–58.
38. Patel KA, Brahmbhatt PK. A comparative study of the RSM
and ANN models for predicting surface roughness in roller
burnishing. Procedia Technol. 2016;23:391–7.
39. Anand K, Shrivastava R, Tamilmannan K, Sathiya P. A
comparative study of artificial neural network and response
surface methodology for optimization of friction welding of
Incoloy 800 H. Acta Metall Sin-EngL. 2015;28(7):892–902.